I worked briefly on a similar idea last year. The problem was to collect parallel data for translation without requiring the annotators to be bilingual. We achieved it by selecting a set of universal pictures (e.g., dog running on grass) and asking the native speakers of Hindi and English to caption it. Since the concepts in the pictures we selected were common across the languages, we were able to collect high quality data.
Paper: https://arxiv.org/abs/2004.11954
Slides: https://madaan.github.io/res/artifacts/pml4dc-practical-data...